This time is different for humanoid robotics because large language models have essentially solved perception, and the combination of safety-certified hardware and sub-$1-per-hour operating costs makes widespread deployment inevitable.

49:08>> yeah well I would say generally it is very easy to make a robot that looks like a person >> and that's why we've seen humanoid for 100 years in one. It's very hard to make a robot that can do useful things in human spaces. >> And we're starting to see that today and that's the difference. So even if it doesn't look exactly like a human, but

49:25maybe a little bit humanoid, but it's doing useful work, >> that's where the impact matters. >> And because of large language models, >> a lot of things have now become free. When these robots look at a table here, >> Yeah. >> and you say, "What's on the table?" It knows that's a phone. It knows this is paper, tea, water. It probably knows how many ounces are in each.

49:47>> Yeah. >> If it were sitting here 3 or 4 years ago, >> it wouldn't actually know >> what was in the world. You would have to program it in a very narrow way. Yeah. >> Yeah. Perception was incredibly difficult. And the fact that perception is all but solved at this point is a really, really huge inflection point. I mean, you know, I said, yes, robots doing useful things, but also people can

50:07now see the future of generality. AI is really enabling that much more broad um you know context awareness for these robots so people can see that this is going to be useful gen generally doing many useful things very soon. >> So there's perception the robot has to understand the world. >> Yep. But then there always seemed to be this blocker with getting the robot out of a very confined narrow task like you

50:32know in a factory >> and I my perception is it was the communication and the training level. Maybe we can unpack that a bit because my understanding was previously you basically had to hardcode the robot if you were going to make a cup of coffee. We have a company I invested in Cafe X and it is a robotic arm. >> Mhm. makes a cup of coffee perfectly

50:55every time, can draft a beer, all that stuff, but it had to be manually coded. Now, the instruction set because of perception, because of language models, having trained on every video on the internet, every coffee recipe that also seems to be for free. Am I wrong or >> not yet? It's actually quite different. So, language models, think of it like it's a it's now becoming kind of a

51:18commodity like the internet. It's available to everybody. It's this amazing rising tide. But these language models are trained off of the entire data on the internet and that data does not exist for robot control. You know what's the example for your robot of all the torqus all the torque commands to every motor given all the sensor input. There's no training set of data. So you have to generate and create that somehow

51:40>> and there's a lot of different approaches and ways people are are going about this. And some of these AI tools again think of AI not as a blackbox but as a big tent of many different very different useful computational tools right in order to control a robot you can do these things by learning from demonstration you can give it you can tellyoperate the robot start to train